Microsoft AI CEO Mustafa Suleyman told employees that a model developed internally outperformed a similar product from OpenAI, one of the software giant’s largest customers and partners.
Suleiman said at a company-wide all-hands meeting on Thursday: “We now have a pre-trained foundational model that is stronger than the one currently powering GPT 5.6 Sol.” He also said: “In fact, we used only one-quarter of the compute power OpenAI used to train its model.”
Pre-training is the first stage in building an AI model, during which the model learns patterns from vast amounts of text, images, and other data. This requires significant and expensive computational resources, so if Microsoft finds more efficient methods, it could reduce costs.
Microsoft announced the model, named MAI-Base-2, earlier on Thursday.
Suleiman said Microsoft privately tested this model against OpenAI’s models, including one nicknamed “Spud,” which powers GPT-5.6. He added that Microsoft’s model performed better in tests measuring AI’s ability to understand and predict patterns in information.
Suleiman stated that these comparisons are based on technical information about OpenAI's models held by Microsoft.
“We also privately and confidentially compared it with OpenAI’s models,” Sulaiman told employees, implying that Microsoft had special access to information about OpenAI’s model performance.
Notably, even in private discussions, Sulaiman chose to emphasize the limitations of OpenAI’s models. Microsoft has invested billions of dollars in OpenAI and holds an equity stake in the startup valued at over $100 billion. Any criticism of OpenAI could undermine the value of this massive investment.
The two companies also maintain a significant partnership, with Microsoft providing OpenAI models to its Azure cloud customers. Microsoft’s own products, such as Copilot, have also consistently used OpenAI models.
However, over the past year or so, the relationship between the two companies has become strained, as Microsoft has increasingly sought to reduce its reliance on OpenAI, including by developing its own models.
Suleiman also stated that Microsoft's new AI model is "12 times more computationally efficient" than the one used by DeepSeek, for example.
Suleiman did not specify which DeepSeek model he was referring to, but a publicly available blog post stated that MAI-Base-2 is significantly more efficient in pretraining than a similar product from this Chinese lab, DeepSeek-V4-Pro-Base.
In March this year, Microsoft formed a new joint team responsible for its Copilot products for both business and consumer markets, and reassigned Sulaiman to the company’s superintelligence division to focus on building AI models.
Microsoft declined to comment, and OpenAI did not respond to requests for comment on Thursday.